fs/coredump.c

   1 #include <linux/slab.h>
   2 #include <linux/file.h>
   3 #include <linux/fdtable.h>
   4 #include <linux/freezer.h>
   5 #include <linux/mm.h>
   6 #include <linux/stat.h>
   7 #include <linux/fcntl.h>
   8 #include <linux/swap.h>
   9 #include <linux/string.h>
  10 #include <linux/init.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/perf_event.h>
  13 #include <linux/highmem.h>
  14 #include <linux/spinlock.h>
  15 #include <linux/key.h>
  16 #include <linux/personality.h>
  17 #include <linux/binfmts.h>
  18 #include <linux/coredump.h>
  19 #include <linux/sched/coredump.h>
  20 #include <linux/utsname.h>
  21 #include <linux/pid_namespace.h>
  22 #include <linux/module.h>
  23 #include <linux/namei.h>
  24 #include <linux/mount.h>
  25 #include <linux/security.h>
  26 #include <linux/syscalls.h>
  27 #include <linux/tsacct_kern.h>
  28 #include <linux/cn_proc.h>
  29 #include <linux/audit.h>
  30 #include <linux/tracehook.h>
  31 #include <linux/kmod.h>
  32 #include <linux/fsnotify.h>
  33 #include <linux/fs_struct.h>
  34 #include <linux/pipe_fs_i.h>
  35 #include <linux/oom.h>
  36 #include <linux/compat.h>
  37 #include <linux/sched.h>
  38 #include <linux/fs.h>
  39 #include <linux/path.h>
  40 #include <linux/timekeeping.h>
  41
  42 #include <linux/uaccess.h>
  43 #include <asm/mmu_context.h>
  44 #include <asm/tlb.h>
  45 #include <asm/exec.h>
  46
  47 #include <trace/events/task.h>
  48 #include "internal.h"
  49
  50 #include <trace/events/sched.h>
  51
  52 int core_uses_pid;
  53 unsigned int core_pipe_limit;
  54 char core_pattern[CORENAME_MAX_SIZE] = "core";
  55 static int core_name_size = CORENAME_MAX_SIZE;
  56
  57 struct core_name {
  58         char *corename;
  59         int used, size;
  60 };
  61
  62 /* The maximal length of core_pattern is also specified in sysctl.c */
  63
  64 static int expand_corename(struct core_name *cn, int size)
  65 {
  66         char *corename = krealloc(cn->corename, size, GFP_KERNEL);
  67
  68         if (!corename)
  69                 return -ENOMEM;
  70
  71         if (size > core_name_size) /* racy but harmless */
  72                 core_name_size = size;
  73
  74         cn->size = ksize(corename);
  75         cn->corename = corename;
  76         return 0;
  77 }
  78
  79 static __printf(2, 0) int cn_vprintf(struct core_name *cn, const char *fmt,
  80                                      va_list arg)
  81 {
  82         int free, need;
  83         va_list arg_copy;
  84
  85 again:
  86         free = cn->size - cn->used;
  87
  88         va_copy(arg_copy, arg);
  89         need = vsnprintf(cn->corename + cn->used, free, fmt, arg_copy);
  90         va_end(arg_copy);
  91
  92         if (need < free) {
  93                 cn->used += need;
  94                 return 0;
  95         }
  96
  97         if (!expand_corename(cn, cn->size + need - free + 1))
  98                 goto again;
  99
 100         return -ENOMEM;
 101 }
 102
 103 static __printf(2, 3) int cn_printf(struct core_name *cn, const char *fmt, ...)
 104 {
 105         va_list arg;
 106         int ret;
 107
 108         va_start(arg, fmt);
 109         ret = cn_vprintf(cn, fmt, arg);
 110         va_end(arg);
 111
 112         return ret;
 113 }
 114
 115 static __printf(2, 3)
 116 int cn_esc_printf(struct core_name *cn, const char *fmt, ...)
 117 {
 118         int cur = cn->used;
 119         va_list arg;
 120         int ret;
 121
 122         va_start(arg, fmt);
 123         ret = cn_vprintf(cn, fmt, arg);
 124         va_end(arg);
 125
 126         if (ret == 0) {
 127                 /*
 128                  * Ensure that this coredump name component can't cause the
 129                  * resulting corefile path to consist of a ".." or ".".
 130                  */
 131                 if ((cn->used - cur == 1 && cn->corename[cur] == '.') ||
 132                                 (cn->used - cur == 2 && cn->corename[cur] == '.'
 133                                 && cn->corename[cur+1] == '.'))
 134                         cn->corename[cur] = '!';
 135
 136                 /*
 137                  * Empty names are fishy and could be used to create a "//" in a
 138                  * corefile name, causing the coredump to happen one directory
 139                  * level too high. Enforce that all components of the core
 140                  * pattern are at least one character long.
 141                  */
 142                 if (cn->used == cur)
 143                         ret = cn_printf(cn, "!");
 144         }
 145
 146         for (; cur < cn->used; ++cur) {
 147                 if (cn->corename[cur] == '/')
 148                         cn->corename[cur] = '!';
 149         }
 150         return ret;
 151 }
 152
 153 static int cn_print_exe_file(struct core_name *cn)
 154 {
 155         struct file *exe_file;
 156         char *pathbuf, *path;
 157         int ret;
 158
 159         exe_file = get_mm_exe_file(current->mm);
 160         if (!exe_file)
 161                 return cn_esc_printf(cn, "%s (path unknown)", current->comm);
 162
 163         pathbuf = kmalloc(PATH_MAX, GFP_TEMPORARY);
 164         if (!pathbuf) {
 165                 ret = -ENOMEM;
 166                 goto put_exe_file;
 167         }
 168
 169         path = file_path(exe_file, pathbuf, PATH_MAX);
 170         if (IS_ERR(path)) {
 171                 ret = PTR_ERR(path);
 172                 goto free_buf;
 173         }
 174
 175         ret = cn_esc_printf(cn, "%s", path);
 176
 177 free_buf:
 178         kfree(pathbuf);
 179 put_exe_file:
 180         fput(exe_file);
 181         return ret;
 182 }
 183
 184 /* format_corename will inspect the pattern parameter, and output a
 185  * name into corename, which must have space for at least
 186  * CORENAME_MAX_SIZE bytes plus one byte for the zero terminator.
 187  */
 188 static int format_corename(struct core_name *cn, struct coredump_params *cprm)
 189 {
 190         const struct cred *cred = current_cred();
 191         const char *pat_ptr = core_pattern;
 192         int ispipe = (*pat_ptr == '|');
 193         int pid_in_pattern = 0;
 194         int err = 0;
 195
 196         cn->used = 0;
 197         cn->corename = NULL;
 198         if (expand_corename(cn, core_name_size))
 199                 return -ENOMEM;
 200         cn->corename[0] = '\0';
 201
 202         if (ispipe)
 203                 ++pat_ptr;
 204
 205         /* Repeat as long as we have more pattern to process and more output
 206            space */
 207         while (*pat_ptr) {
 208                 if (*pat_ptr != '%') {
 209                         err = cn_printf(cn, "%c", *pat_ptr++);
 210                 } else {
 211                         switch (*++pat_ptr) {
 212                         /* single % at the end, drop that */
 213                         case 0:
 214                                 goto out;
 215                         /* Double percent, output one percent */
 216                         case '%':
 217                                 err = cn_printf(cn, "%c", '%');
 218                                 break;
 219                         /* pid */
 220                         case 'p':
 221                                 pid_in_pattern = 1;
 222                                 err = cn_printf(cn, "%d",
 223                                               task_tgid_vnr(current));
 224                                 break;
 225                         /* global pid */
 226                         case 'P':
 227                                 err = cn_printf(cn, "%d",
 228                                               task_tgid_nr(current));
 229                                 break;
 230                         case 'i':
 231                                 err = cn_printf(cn, "%d",
 232                                               task_pid_vnr(current));
 233                                 break;
 234                         case 'I':
 235                                 err = cn_printf(cn, "%d",
 236                                               task_pid_nr(current));
 237                                 break;
 238                         /* uid */
 239                         case 'u':
 240                                 err = cn_printf(cn, "%u",
 241                                                 from_kuid(&init_user_ns,
 242                                                           cred->uid));
 243                                 break;
 244                         /* gid */
 245                         case 'g':
 246                                 err = cn_printf(cn, "%u",
 247                                                 from_kgid(&init_user_ns,
 248                                                           cred->gid));
 249                                 break;
 250                         case 'd':
 251                                 err = cn_printf(cn, "%d",
 252                                         __get_dumpable(cprm->mm_flags));
 253                                 break;
 254                         /* signal that caused the coredump */
 255                         case 's':
 256                                 err = cn_printf(cn, "%d",
 257                                                 cprm->siginfo->si_signo);
 258                                 break;
 259                         /* UNIX time of coredump */
 260                         case 't': {
 261                                 time64_t time;
 262
 263                                 time = ktime_get_real_seconds();
 264                                 err = cn_printf(cn, "%lld", time);
 265                                 break;
 266                         }
 267                         /* hostname */
 268                         case 'h':
 269                                 down_read(&uts_sem);
 270                                 err = cn_esc_printf(cn, "%s",
 271                                               utsname()->nodename);
 272                                 up_read(&uts_sem);
 273                                 break;
 274                         /* executable */
 275                         case 'e':
 276                                 err = cn_esc_printf(cn, "%s", current->comm);
 277                                 break;
 278                         case 'E':
 279                                 err = cn_print_exe_file(cn);
 280                                 break;
 281                         /* core limit size */
 282                         case 'c':
 283                                 err = cn_printf(cn, "%lu",
 284                                               rlimit(RLIMIT_CORE));
 285                                 break;
 286                         default:
 287                                 break;
 288                         }
 289                         ++pat_ptr;
 290                 }
 291
 292                 if (err)
 293                         return err;
 294         }
 295
 296 out:
 297         /* Backward compatibility with core_uses_pid:
 298          *
 299          * If core_pattern does not include a %p (as is the default)
 300          * and core_uses_pid is set, then .%pid will be appended to
 301          * the filename. Do not do this for piped commands. */
 302         if (!ispipe && !pid_in_pattern && core_uses_pid) {
 303                 err = cn_printf(cn, ".%d", task_tgid_vnr(current));
 304                 if (err)
 305                         return err;
 306         }
 307         return ispipe;
 308 }
 309
 310 static int zap_process(struct task_struct *start, int exit_code, int flags)
 311 {
 312         struct task_struct *t;
 313         int nr = 0;
 314
 315         /* ignore all signals except SIGKILL, see prepare_signal() */
 316         start->signal->flags = SIGNAL_GROUP_COREDUMP | flags;
 317         start->signal->group_exit_code = exit_code;
 318         start->signal->group_stop_count = 0;
 319
 320         for_each_thread(start, t) {
 321                 task_clear_jobctl_pending(t, JOBCTL_PENDING_MASK);
 322                 if (t != current && t->mm) {
 323                         sigaddset(&t->pending.signal, SIGKILL);
 324                         signal_wake_up(t, 1);
 325                         nr++;
 326                 }
 327         }
 328
 329         return nr;
 330 }
 331
 332 static int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
 333                         struct core_state *core_state, int exit_code)
 334 {
 335         struct task_struct *g, *p;
 336         unsigned long flags;
 337         int nr = -EAGAIN;
 338
 339         spin_lock_irq(&tsk->sighand->siglock);
 340         if (!signal_group_exit(tsk->signal)) {
 341                 mm->core_state = core_state;
 342                 tsk->signal->group_exit_task = tsk;
 343                 nr = zap_process(tsk, exit_code, 0);
 344                 clear_tsk_thread_flag(tsk, TIF_SIGPENDING);
 345         }
 346         spin_unlock_irq(&tsk->sighand->siglock);
 347         if (unlikely(nr < 0))
 348                 return nr;
 349
 350         tsk->flags |= PF_DUMPCORE;
 351         if (atomic_read(&mm->mm_users) == nr + 1)
 352                 goto done;
 353         /*
 354          * We should find and kill all tasks which use this mm, and we should
 355          * count them correctly into ->nr_threads. We don't take tasklist
 356          * lock, but this is safe wrt:
 357          *
 358          * fork:
 359          *      None of sub-threads can fork after zap_process(leader). All
 360          *      processes which were created before this point should be
 361          *      visible to zap_threads() because copy_process() adds the new
 362          *      process to the tail of init_task.tasks list, and lock/unlock
 363          *      of ->siglock provides a memory barrier.
 364          *
 365          * do_exit:
 366          *      The caller holds mm->mmap_sem. This means that the task which
 367          *      uses this mm can't pass exit_mm(), so it can't exit or clear
 368          *      its ->mm.
 369          *
 370          * de_thread:
 371          *      It does list_replace_rcu(&leader->tasks, &current->tasks),
 372          *      we must see either old or new leader, this does not matter.
 373          *      However, it can change p->sighand, so lock_task_sighand(p)
 374          *      must be used. Since p->mm != NULL and we hold ->mmap_sem
 375          *      it can't fail.
 376          *
 377          *      Note also that "g" can be the old leader with ->mm == NULL
 378          *      and already unhashed and thus removed from ->thread_group.
 379          *      This is OK, __unhash_process()->list_del_rcu() does not
 380          *      clear the ->next pointer, we will find the new leader via
 381          *      next_thread().
 382          */
 383         rcu_read_lock();
 384         for_each_process(g) {
 385                 if (g == tsk->group_leader)
 386                         continue;
 387                 if (g->flags & PF_KTHREAD)
 388                         continue;
 389
 390                 for_each_thread(g, p) {
 391                         if (unlikely(!p->mm))
 392                                 continue;
 393                         if (unlikely(p->mm == mm)) {
 394                                 lock_task_sighand(p, &flags);
 395                                 nr += zap_process(p, exit_code,
 396                                                         SIGNAL_GROUP_EXIT);
 397                                 unlock_task_sighand(p, &flags);
 398                         }
 399                         break;
 400                 }
 401         }
 402         rcu_read_unlock();
 403 done:
 404         atomic_set(&core_state->nr_threads, nr);
 405         return nr;
 406 }
 407
 408 static int coredump_wait(int exit_code, struct core_state *core_state)
 409 {
 410         struct task_struct *tsk = current;
 411         struct mm_struct *mm = tsk->mm;
 412         int core_waiters = -EBUSY;
 413
 414         init_completion(&core_state->startup);
 415         core_state->dumper.task = tsk;
 416         core_state->dumper.next = NULL;
 417
 418         if (down_write_killable(&mm->mmap_sem))
 419                 return -EINTR;
 420
 421         if (!mm->core_state)
 422                 core_waiters = zap_threads(tsk, mm, core_state, exit_code);
 423         up_write(&mm->mmap_sem);
 424
 425         if (core_waiters > 0) {
 426                 struct core_thread *ptr;
 427
 428                 freezer_do_not_count();
 429                 wait_for_completion(&core_state->startup);
 430                 freezer_count();
 431                 /*
 432                  * Wait for all the threads to become inactive, so that
 433                  * all the thread context (extended register state, like
 434                  * fpu etc) gets copied to the memory.
 435                  */
 436                 ptr = core_state->dumper.next;
 437                 while (ptr != NULL) {
 438                         wait_task_inactive(ptr->task, 0);
 439                         ptr = ptr->next;
 440                 }
 441         }
 442
 443         return core_waiters;
 444 }
 445
 446 static void coredump_finish(struct mm_struct *mm, bool core_dumped)
 447 {
 448         struct core_thread *curr, *next;
 449         struct task_struct *task;
 450
 451         spin_lock_irq(&current->sighand->siglock);
 452         if (core_dumped && !__fatal_signal_pending(current))
 453                 current->signal->group_exit_code |= 0x80;
 454         current->signal->group_exit_task = NULL;
 455         current->signal->flags = SIGNAL_GROUP_EXIT;
 456         spin_unlock_irq(&current->sighand->siglock);
 457
 458         next = mm->core_state->dumper.next;
 459         while ((curr = next) != NULL) {
 460                 next = curr->next;
 461                 task = curr->task;
 462                 /*
 463                  * see exit_mm(), curr->task must not see
 464                  * ->task == NULL before we read ->next.
 465                  */
 466                 smp_mb();
 467                 curr->task = NULL;
 468                 wake_up_process(task);
 469         }
 470
 471         mm->core_state = NULL;
 472 }
 473
 474 static bool dump_interrupted(void)
 475 {
 476         /*
 477          * SIGKILL or freezing() interrupt the coredumping. Perhaps we
 478          * can do try_to_freeze() and check __fatal_signal_pending(),
 479          * but then we need to teach dump_write() to restart and clear
 480          * TIF_SIGPENDING.
 481          */
 482         return signal_pending(current);
 483 }
 484
 485 static void wait_for_dump_helpers(struct file *file)
 486 {
 487         struct pipe_inode_info *pipe = file->private_data;
 488
 489         pipe_lock(pipe);
 490         pipe->readers++;
 491         pipe->writers--;
 492         wake_up_interruptible_sync(&pipe->wait);
 493         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 494         pipe_unlock(pipe);
 495
 496         /*
 497          * We actually want wait_event_freezable() but then we need
 498          * to clear TIF_SIGPENDING and improve dump_interrupted().
 499          */
 500         wait_event_interruptible(pipe->wait, pipe->readers == 1);
 501
 502         pipe_lock(pipe);
 503         pipe->readers--;
 504         pipe->writers++;
 505         pipe_unlock(pipe);
 506 }
 507
 508 /*
 509  * umh_pipe_setup
 510  * helper function to customize the process used
 511  * to collect the core in userspace.  Specifically
 512  * it sets up a pipe and installs it as fd 0 (stdin)
 513  * for the process.  Returns 0 on success, or
 514  * PTR_ERR on failure.
 515  * Note that it also sets the core limit to 1.  This
 516  * is a special value that we use to trap recursive
 517  * core dumps
 518  */
 519 static int umh_pipe_setup(struct subprocess_info *info, struct cred *new)
 520 {
 521         struct file *files[2];
 522         struct coredump_params *cp = (struct coredump_params *)info->data;
 523         int err = create_pipe_files(files, 0);
 524         if (err)
 525                 return err;
 526
 527         cp->file = files[1];
 528
 529         err = replace_fd(0, files[0], 0);
 530         fput(files[0]);
 531         /* and disallow core files too */
 532         current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
 533
 534         return err;
 535 }
 536
 537 void do_coredump(const siginfo_t *siginfo)
 538 {
 539         struct core_state core_state;
 540         struct core_name cn;
 541         struct mm_struct *mm = current->mm;
 542         struct linux_binfmt * binfmt;
 543         const struct cred *old_cred;
 544         struct cred *cred;
 545         int retval = 0;
 546         int ispipe;
 547         struct files_struct *displaced;
 548         /* require nonrelative corefile path and be extra careful */
 549         bool need_suid_safe = false;
 550         bool core_dumped = false;
 551         static atomic_t core_dump_count = ATOMIC_INIT(0);
 552         struct coredump_params cprm = {
 553                 .siginfo = siginfo,
 554                 .regs = signal_pt_regs(),
 555                 .limit = rlimit(RLIMIT_CORE),
 556                 /*
 557                  * We must use the same mm->flags while dumping core to avoid
 558                  * inconsistency of bit flags, since this flag is not protected
 559                  * by any locks.
 560                  */
 561                 .mm_flags = mm->flags,
 562         };
 563
 564         audit_core_dumps(siginfo->si_signo);
 565
 566         binfmt = mm->binfmt;
 567         if (!binfmt || !binfmt->core_dump)
 568                 goto fail;
 569         if (!__get_dumpable(cprm.mm_flags))
 570                 goto fail;
 571
 572         cred = prepare_creds();
 573         if (!cred)
 574                 goto fail;
 575         /*
 576          * We cannot trust fsuid as being the "true" uid of the process
 577          * nor do we know its entire history. We only know it was tainted
 578          * so we dump it as root in mode 2, and only into a controlled
 579          * environment (pipe handler or fully qualified path).
 580          */
 581         if (__get_dumpable(cprm.mm_flags) == SUID_DUMP_ROOT) {
 582                 /* Setuid core dump mode */
 583                 cred->fsuid = GLOBAL_ROOT_UID;  /* Dump root private */
 584                 need_suid_safe = true;
 585         }
 586
 587         retval = coredump_wait(siginfo->si_signo, &core_state);
 588         if (retval < 0)
 589                 goto fail_creds;
 590
 591         old_cred = override_creds(cred);
 592
 593         ispipe = format_corename(&cn, &cprm);
 594
 595         if (ispipe) {
 596                 int dump_count;
 597                 char **helper_argv;
 598                 struct subprocess_info *sub_info;
 599
 600                 if (ispipe < 0) {
 601                         printk(KERN_WARNING "format_corename failed\n");
 602                         printk(KERN_WARNING "Aborting core\n");
 603                         goto fail_unlock;
 604                 }
 605
 606                 if (cprm.limit == 1) {
 607                         /* See umh_pipe_setup() which sets RLIMIT_CORE = 1.
 608                          *
 609                          * Normally core limits are irrelevant to pipes, since
 610                          * we're not writing to the file system, but we use
 611                          * cprm.limit of 1 here as a special value, this is a
 612                          * consistent way to catch recursive crashes.
 613                          * We can still crash if the core_pattern binary sets
 614                          * RLIM_CORE = !1, but it runs as root, and can do
 615                          * lots of stupid things.
 616                          *
 617                          * Note that we use task_tgid_vnr here to grab the pid
 618                          * of the process group leader.  That way we get the
 619                          * right pid if a thread in a multi-threaded
 620                          * core_pattern process dies.
 621                          */
 622                         printk(KERN_WARNING
 623                                 "Process %d(%s) has RLIMIT_CORE set to 1\n",
 624                                 task_tgid_vnr(current), current->comm);
 625                         printk(KERN_WARNING "Aborting core\n");
 626                         goto fail_unlock;
 627                 }
 628                 cprm.limit = RLIM_INFINITY;
 629
 630                 dump_count = atomic_inc_return(&core_dump_count);
 631                 if (core_pipe_limit && (core_pipe_limit < dump_count)) {
 632                         printk(KERN_WARNING "Pid %d(%s) over core_pipe_limit\n",
 633                                task_tgid_vnr(current), current->comm);
 634                         printk(KERN_WARNING "Skipping core dump\n");
 635                         goto fail_dropcount;
 636                 }
 637
 638                 helper_argv = argv_split(GFP_KERNEL, cn.corename, NULL);
 639                 if (!helper_argv) {
 640                         printk(KERN_WARNING "%s failed to allocate memory\n",
 641                                __func__);
 642                         goto fail_dropcount;
 643                 }
 644
 645                 retval = -ENOMEM;
 646                 sub_info = call_usermodehelper_setup(helper_argv[0],
 647                                                 helper_argv, NULL, GFP_KERNEL,
 648                                                 umh_pipe_setup, NULL, &cprm);
 649                 if (sub_info)
 650                         retval = call_usermodehelper_exec(sub_info,
 651                                                           UMH_WAIT_EXEC);
 652
 653                 argv_free(helper_argv);
 654                 if (retval) {
 655                         printk(KERN_INFO "Core dump to |%s pipe failed\n",
 656                                cn.corename);
 657                         goto close_fail;
 658                 }
 659         } else {
 660                 struct inode *inode;
 661                 int open_flags = O_CREAT | O_RDWR | O_NOFOLLOW |
 662                                  O_LARGEFILE | O_EXCL;
 663
 664                 if (cprm.limit < binfmt->min_coredump)
 665                         goto fail_unlock;
 666
 667                 if (need_suid_safe && cn.corename[0] != '/') {
 668                         printk(KERN_WARNING "Pid %d(%s) can only dump core "\
 669                                 "to fully qualified path!\n",
 670                                 task_tgid_vnr(current), current->comm);
 671                         printk(KERN_WARNING "Skipping core dump\n");
 672                         goto fail_unlock;
 673                 }
 674
 675                 /*
 676                  * Unlink the file if it exists unless this is a SUID
 677                  * binary - in that case, we're running around with root
 678                  * privs and don't want to unlink another user's coredump.
 679                  */
 680                 if (!need_suid_safe) {
 681                         mm_segment_t old_fs;
 682
 683                         old_fs = get_fs();
 684                         set_fs(KERNEL_DS);
 685                         /*
 686                          * If it doesn't exist, that's fine. If there's some
 687                          * other problem, we'll catch it at the filp_open().
 688                          */
 689                         (void) sys_unlink((const char __user *)cn.corename);
 690                         set_fs(old_fs);
 691                 }
 692
 693                 /*
 694                  * There is a race between unlinking and creating the
 695                  * file, but if that causes an EEXIST here, that's
 696                  * fine - another process raced with us while creating
 697                  * the corefile, and the other process won. To userspace,
 698                  * what matters is that at least one of the two processes
 699                  * writes its coredump successfully, not which one.
 700                  */
 701                 if (need_suid_safe) {
 702                         /*
 703                          * Using user namespaces, normal user tasks can change
 704                          * their current->fs->root to point to arbitrary
 705                          * directories. Since the intention of the "only dump
 706                          * with a fully qualified path" rule is to control where
 707                          * coredumps may be placed using root privileges,
 708                          * current->fs->root must not be used. Instead, use the
 709                          * root directory of init_task.
 710                          */
 711                         struct path root;
 712
 713                         task_lock(&init_task);
 714                         get_fs_root(init_task.fs, &root);
 715                         task_unlock(&init_task);
 716                         cprm.file = file_open_root(root.dentry, root.mnt,
 717                                 cn.corename, open_flags, 0600);
 718                         path_put(&root);
 719                 } else {
 720                         cprm.file = filp_open(cn.corename, open_flags, 0600);
 721                 }
 722                 if (IS_ERR(cprm.file))
 723                         goto fail_unlock;
 724
 725                 inode = file_inode(cprm.file);
 726                 if (inode->i_nlink > 1)
 727                         goto close_fail;
 728                 if (d_unhashed(cprm.file->f_path.dentry))
 729                         goto close_fail;
 730                 /*
 731                  * AK: actually i see no reason to not allow this for named
 732                  * pipes etc, but keep the previous behaviour for now.
 733                  */
 734                 if (!S_ISREG(inode->i_mode))
 735                         goto close_fail;
 736                 /*
 737                  * Don't dump core if the filesystem changed owner or mode
 738                  * of the file during file creation. This is an issue when
 739                  * a process dumps core while its cwd is e.g. on a vfat
 740                  * filesystem.
 741                  */
 742                 if (!uid_eq(inode->i_uid, current_fsuid()))
 743                         goto close_fail;
 744                 if ((inode->i_mode & 0677) != 0600)
 745                         goto close_fail;
 746                 if (!(cprm.file->f_mode & FMODE_CAN_WRITE))
 747                         goto close_fail;
 748                 if (do_truncate(cprm.file->f_path.dentry, 0, 0, cprm.file))
 749                         goto close_fail;
 750         }
 751
 752         /* get us an unshared descriptor table; almost always a no-op */
 753         retval = unshare_files(&displaced);
 754         if (retval)
 755                 goto close_fail;
 756         if (displaced)
 757                 put_files_struct(displaced);
 758         if (!dump_interrupted()) {
 759                 file_start_write(cprm.file);
 760                 core_dumped = binfmt->core_dump(&cprm);
 761                 file_end_write(cprm.file);
 762         }
 763         if (ispipe && core_pipe_limit)
 764                 wait_for_dump_helpers(cprm.file);
 765 close_fail:
 766         if (cprm.file)
 767                 filp_close(cprm.file, NULL);
 768 fail_dropcount:
 769         if (ispipe)
 770                 atomic_dec(&core_dump_count);
 771 fail_unlock:
 772         kfree(cn.corename);
 773         coredump_finish(mm, core_dumped);
 774         revert_creds(old_cred);
 775 fail_creds:
 776         put_cred(cred);
 777 fail:
 778         return;
 779 }
 780
 781 /*
 782  * Core dumping helper functions.  These are the only things you should
 783  * do on a core-file: use only these functions to write out all the
 784  * necessary info.
 785  */
 786 int dump_emit(struct coredump_params *cprm, const void *addr, int nr)
 787 {
 788         struct file *file = cprm->file;
 789         loff_t pos = file->f_pos;
 790         ssize_t n;
 791         if (cprm->written + nr > cprm->limit)
 792                 return 0;
 793         while (nr) {
 794                 if (dump_interrupted())
 795                         return 0;
 796                 n = __kernel_write(file, addr, nr, &pos);
 797                 if (n <= 0)
 798                         return 0;
 799                 file->f_pos = pos;
 800                 cprm->written += n;
 801                 cprm->pos += n;
 802                 nr -= n;
 803         }
 804         return 1;
 805 }
 806 EXPORT_SYMBOL(dump_emit);
 807
 808 int dump_skip(struct coredump_params *cprm, size_t nr)
 809 {
 810         static char zeroes[PAGE_SIZE];
 811         struct file *file = cprm->file;
 812         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
 813                 if (dump_interrupted() ||
 814                     file->f_op->llseek(file, nr, SEEK_CUR) < 0)
 815                         return 0;
 816                 cprm->pos += nr;
 817                 return 1;
 818         } else {
 819                 while (nr > PAGE_SIZE) {
 820                         if (!dump_emit(cprm, zeroes, PAGE_SIZE))
 821                                 return 0;
 822                         nr -= PAGE_SIZE;
 823                 }
 824                 return dump_emit(cprm, zeroes, nr);
 825         }
 826 }
 827 EXPORT_SYMBOL(dump_skip);
 828
 829 int dump_align(struct coredump_params *cprm, int align)
 830 {
 831         unsigned mod = cprm->pos & (align - 1);
 832         if (align & (align - 1))
 833                 return 0;
 834         return mod ? dump_skip(cprm, align - mod) : 1;
 835 }
 836 EXPORT_SYMBOL(dump_align);
 837
 838 /*
 839  * Ensures that file size is big enough to contain the current file
 840  * postion. This prevents gdb from complaining about a truncated file
 841  * if the last "write" to the file was dump_skip.
 842  */
 843 void dump_truncate(struct coredump_params *cprm)
 844 {
 845         struct file *file = cprm->file;
 846         loff_t offset;
 847
 848         if (file->f_op->llseek && file->f_op->llseek != no_llseek) {
 849                 offset = file->f_op->llseek(file, 0, SEEK_CUR);
 850                 if (i_size_read(file->f_mapping->host) < offset)
 851                         do_truncate(file->f_path.dentry, offset, 0, file);
 852         }
 853 }
 854 EXPORT_SYMBOL(dump_truncate);